In a report dated 29 November 2011, The Guardian newspaper asked ‘[w]hy did NATO forces kill two dozen Pakistani soldiers at a border post in the Mohmand region, some 300 yards across the frontier from Afghanistan early on Saturday morning?’Footnote 1 Having reflected upon differing explanations for the event, the report asserted ‘[t]here is a very simple explanation of what happened, the US military makes deadly mistakes all the time, and for all its technological wizardry and tremendous firepower, it has very little intelligence on the ground’. Reportedly, in 2010 ‘a U.S. military investigation … harshly criticized a Nevada-based Air Force drone crew and American ground commanders in Afghanistan for misidentifying civilians as insurgents during a U.S. Army Special Forces operation in Oruzgan province in February, resulting in the deaths of as many as 23 civilians’.Footnote 2
From one kind of ‘military operations from a distance’, or remote attack as we shall call the phenomenon, let us move to another, namely cyber operations. Military use of cyber operationsFootnote 3 occurred on 27 and 28 April 2007 when an apparently coordinated sequence of denial-of-service operations affected websites in Estonia during a dispute between that country and Russia. Ping requests were followed by malformed Web queries to governmental and media websites. From 30 April until 18 May 2007, distributed operations aimed at producing a denial of service from targeted websites (distributed denial of service or DDoS) followed. Careful timing of cyber operations maximized their effectiveness, and the affected sites became temporarily inaccessible. It appeared that botnets were being employed and a precise impact was the evident result.Footnote 4 Some Estonian websites were defaced by so-called patriotic hackers, but it was never formally determined which state, if any, was responsible.Footnote 5 Then, in 2008, cyber operations were undertaken against Georgia during its armed conflict with Russia.
The 2010 Stuxnet operation against Iran was, perhaps, one of the more militarily significant cyber operations. Stuxnet is an integrated set of components that were used to undertake computer network attacks. Using, in part, a worm as its delivery mechanism, Stuxnet inserts itself onto disconnected networks, for example through the use of thumb drives or CD-ROMs. It searches for a specified manufacturer's model of computer control facility – in the case of the Iranian attack a control system manufactured by Siemens – finds and places itself on a relevant node and undertakes pre-planned activity. During the July 2010 operation, malware reportedly attacked centrifuges evidently associated with the Iranian nuclear programme and, it appears, caused damage.Footnote 6 While the defacement of websites as exemplified in the Estonian operations would not seem to amount to an attack in the in bello sense,Footnote 7 it is likely that the Stuxnet attack would be regarded at law as such an attack because of the damage reportedly caused to the centrifuges.
The use, during armed conflicts, of these cyber techniques to prosecute attacks, that is to cause death, injury, damage or destruction, or the employment of remotely pilotedFootnote 8 or, in the future, autonomous unmanned platforms to undertake attacks constitutes what, for the purposes of this article, we shall describe as ‘remote attack’. Such attacks are remote in the sense that the operator of the remotely piloted vehicle or the initiator of the autonomous mission or of the cyber attack is liable to be located at a considerable distance from the scene of the injury or destruction wrought by the attack. The purpose of the present article is to consider whether the remote conduct of attacks using such techniques during armed conflicts raises legal concerns. The author's starting point is that cyber attacks during armed conflict, namely military operations in which cyber means are employed to inflict death, injury, damage or destruction on an adverse party to the conflict, are regulated by the law of armed conflict and thus, for states party to the Protocol Additional to the Geneva Conventions of 12 August 1949 and relating to the Protection of Victims of International Armed Conflicts (API),Footnote 9 are subject to the rules in Articles 48 to 67 of that treaty.Footnote 10 For states that are not party to API, the customary principles and rules – most notably the customary principle of distinction and the customary rules of discrimination, of proportionality, and of precautions in attack – will apply.Footnote 11 Similarly, it seems to be generally accepted that the same body of law regulates attacks using unmanned platforms, that is aircraft, ground vehicles, ships or other marine craft that do not carry crew personnel and that are either controlled by an operator who is located remotely from the relevant platform or that employ autonomous guidance and attack technology.Footnote 12 We will discuss these issues primarily by reference to the air domain and will call such operator-controlled vehicles ‘remotely piloted vehicles’, while references to autonomy will be applied to platforms that make attack decisions without the supervision of a human being. In relation to both such methods of attack, the question to be discussed is therefore whether the absence of the person who is undertaking the attack from the location of its operative effect raises legal concerns.
We shall start by considering attacks using remotely piloted platforms. We will then briefly outline the issues in relation to precautions in attack posed by the use of autonomous attack technologies. In the third section of the article we will summarize how the targeting rules in API can be applied to cyber attacks. Then, in the fourth section, we will analyse where the remoteness challenge sits. In the fifth section we will discuss where liability may rest for these differing classes of attack. In the final substantive section we will ask whether these new technologies represent a qualitative change in the conduct of warfare or a further development in a well-established evolutionary process, essentially posing the question whether what we are discussing is really anything substantively new. We will then seek to draw conclusions.
Remotely piloted vehicles and the law
The remoteness of the controller from the attack does not, per se, exclude the application of targeting law to such activities. The legal principle of distinction,Footnote 13 the prohibition of indiscriminate attacks,Footnote 14 the precautions rules, and the more detailed provisions requiring the protection of specific persons and objectsFootnote 15 will all apply to such operations. The controller of a Predator or Reaper Unmanned Aerial Vehicle (UAV), although located some thousands of miles from the scene of the attack, bases his attack decisions on the information derived from sensors and other sources and is as constrained by the targeting rules, including the rules as to precautions in attack, as any other military operator in the battle space, including a pilot of a manned aircraft.
Accordingly, the UAV operator must take constant care to spare civilians and civilian objects when undertaking military operations in general;Footnote 16 he must do everything practicable or practically possibleFootnote 17 to ‘verify that the objectives to be attacked are neither civilians nor civilian objects and are not subject to special protection but are military objectives … and that it is not prohibited … to attack them’; he must take all practicable or practically possible precautions in the choice of means and methods of attack with a view to avoiding, and in any event to minimizing, incidental loss of civilian life, injury to civilians and damage to civilian objects;Footnote 18 he must ‘refrain from deciding to launch any attack which may be expected’ to cause disproportionate incidental civilian injury and/or damage;Footnote 19 he must cancel or suspend the attack if it becomes clear that its objective is not a military objective, that its objective is subject to special protection or that the attack may be expected to cause disproportionate incidental civilian injury or damage;Footnote 20 he must ensure that an effective advance warning is given if civilians may be affected by the attack unless circumstances do not permit;Footnote 21 and he must ensure that ‘when a choice is possible between several military objectives for obtaining a similar military advantage, the objective that is selected is the objective ‘the attack on which may be expected to cause the least danger to civilian lives and to civilian objects’.Footnote 22 These precautionary rules bind parties to API as a matter of treaty law and, as we noted above, are largely customary and thus bind all states. It follows from this analysis that the precautionary duties of a controller of an armed UAV are just as exacting as those imposed on the pilot of a manned aircraft. The law does not reduce these duties because of the absence of a person from the cockpit.Footnote 23
Autonomous attack and the law
The word ‘autonomy’ is taken for the purposes of the present discussion to refer to autonomous attack decision-making undertaken by, for example, algorithm-based technology on board an unmanned platform such as an aerial vehicle.Footnote 24 The technology may, for example, be programmed to detect points of recognition of particular military objects, such as a tank, artillery piece or armoured personnel carrier. If the technology adequately distinguishes between such military objects and civilian objects, it would seem that the requirement in Article 57(2)(a)(i) of APIFootnote 25 may be complied with, provided it can properly be said that ‘everything feasible’ is being done to accomplish the required distinction. In the light of the United Kingdom (UK) interpretative statement cited above,Footnote 26 military considerations may be taken into account in order to determine that which is practically possible and thus required as a feasible precaution. An argument that the absence of a human being from the autonomous aspect of the decision-making process renders the performance of these precautionary duties impractical and that they are therefore to be regarded as militarily non-feasible would, in the author's view, be unsatisfactory, not least because alternative methods of undertaking such attacks would permit of the taking of such precautions. The better view must therefore be that the full set of precautionary measures set out in Article 57 of API and summarized in the previous section of this article must be complied with in relation to autonomous attacks.
While compliance with Article 57(2)(a)(i) of API may be achievable as discussed in the previous paragraph,Footnote 27 things get somewhat more difficult when we consider the evaluative rules of precaution. These further precautionary duties, listed in the previous section and which do not require repetition here, generate the challenging question of whether technology is capable of mechanizing essentially evaluative judgements. These include the assessment of whether the chosen means and method for undertaking the planned attack will in fact minimize injury to civilians and damage to civilian objects and whether the injury to civilians and the damage to civilian objects that may be expected to result from the attack of a given class of military objective on a specified occasion will be excessive in relation to the anticipated military advantage. The statement by the UK and other states on ratification of API, to the effect that military advantage is intended to refer to that accruing from the attack considered as a whole,Footnote 28 suggests that the proportionality assessment should be applied to something more that an individual engagement of a single object.Footnote 29
Nevertheless, a means or method of warfareFootnote 30 is likely to prove legally unacceptable if it precludes the taking of these legally required evaluative precautions. Autonomous attack methods will not, however, necessarily preclude the taking of these precautions. Thus, planners and operational decision-makers contemplating the mounting of an autonomous mission are likely to be in a position to review relevant pattern-of-life data relating to the planned area of search. They will review that data in order to assess, before the commencement of the autonomous mission, the civilian death, injury, and damage that may be expected as a result of an attack of the planned class of military objective in that area during the planned period of search using the weapons loaded onto the platform. The military advantage to be anticipated from the successful attack of an object that the algorithm technology is programmed to recognize will be known at the planning stage, so, depending on the pattern of life in the relevant area, it may be possible to comply with the evaluative precautionary rules at the mission planning stage thus rendering the use of autonomous attack technology potentially lawful. This is most likely to be the case if the planned area of search is remote from civilians and civilian objects; areas of desert, remote steppe lands, and remote maritime areas would seem to be examples. It may also be the case if, for whatever reason, pattern-of-life data clearly show that civilians will remain absent from a less remote area at the time of the planned search.
If, by contrast, judgements as to the minimization of civilian death, injury, and damage and as to the proportionality of attacks cannot be made at the sortie planning stage, for example because of the congested urban nature of the area of search or because for whatever reason civilian death, injury, and damage cannot be predicted with acceptable assurance in advance of the mission, it follows that the evaluative precautions cannot be undertaken with the consequence that a decision to undertake an autonomous mission in such circumstances would breach Article 57.
The focus in this discussion is on autonomous attacks targeting inherently military objects with characteristics that facilitate mechanical recognition. So far as is known, technology is not currently available to support the autonomous distinguishing of military personnel from civilians. Only when autonomous attack technology can make those distinctions to an acceptable degree of reliability, and only when, having so distinguished, the technology enables the evaluative decisions referred to above to be made in the context of attacks that target persons will there be any basis for a discussion of autonomous attack of individuals. The author is not aware of any such system yet having been fielded, and therefore concludes that autonomous attack of personnel can, for the time being at least, be excluded on the ground that the rules as to precautions in attack cannot be complied with.Footnote 31
Cyber attacks and the law
The computer age has brought into existence another environment in which hostilities can be conducted.Footnote 32 The dependence of modern societies and of their armed forces on computer systems renders such systems prime objects of attack, or a choice medium through which to target some linked object or person.Footnote 33 Events in Estonia in 2007,Footnote 34 in Georgia in 2008Footnote 35 and in Iran in 2010Footnote 36 indicate that the offensive use of cyber operations will be an increasingly important aspect of warfare in coming decades. Cyber operations can be taken to be military operations in which one computer is used either to target another or to use that other computer as the conduit through which injury or damage is caused to an opposing party to the conflict. The use of any instrument, including a computer, to cause death, injury, damage or destruction to another party to an armed conflict will cause that instrument, or computer, to become a weapon or means of warfare.Footnote 37 The damage or injury may be caused to the users of the targeted computer system or the targeted system itself may be damaged; in either case causing the cyber operation to be regarded as a cyber attack. The critical issue for the purposes of the present article is, however, that the operation may be initiated a considerable distance in both space and time from the place and time, where and when, the damaging consequences are intended to occur. This notion of remoteness of the operator from the consequences of his or her activity is compounded by the difficulty that is likely to be encountered in determining, and then being able to demonstrate, first, who undertook the cyber operation in question, second, on behalf of which state or organization, if any, the operation was undertaken, and, third, its purpose.
A relevant legal issue arises from the difficulty that the planner and decision-maker are likely to have in evaluating in advance the expected results of a planned cyber attack. In order to make any sensible assessment of the legitimacy of the planned attack they will need to know enough about the cyber linkages between the sending computer and the targeted computer to be sufficiently assured that the attack will in fact engage the intended target. Secondly, they will also need to know enough about the characteristics of the particular cyber capability that is being used to undertake the attack to be assured that it will engage the target in the intended way. Thirdly, they will need to know enough about the targeted computer system, its dependencies, and associated networks to be able to assess the proportionality of the planned attack. Finally, if the cyber capability to be used in the attack is liable to affect other networks as it travels to the targeted system, the expected effects on those other networks will need to be assessed as, to the extent that those networks do not themselves consist of military objectives, damage to them, and consequential damage or injury to their users will have to be factored into the proportionality assessment that is made in advance of the decision to mount the cyber attack.
Mapping the targeted system, its dependencies, and the intervening linkages in this way is likely to be a challenging task. Undertaking that mapping in a covert way is likely to be even more difficult. To maintain that operational security by failing to undertake any assessment of the proportionality of the planned attack is likely to breach Article 57 for the same reasons as were noted in the previous section.
Where the remoteness challenge sits
What emerges from the analysis, however, is that the distance in time and space does not of itself render the attack unlawful. At the root of the problem is the effect that this remoteness has on the ability of planners and decision-makers to undertake required precautions and to obtain information to support a sensible evaluation of the lawfulness of the planned attack. To put the matter simply, it is only when the technological advances that enable remote attack, be it cyber, autonomous or remotely piloted, are matched by the technological capability to inform the standard precautions the law requires in relation to all attacks that the use of such remote attack capabilities becomes lawful. This has been broadly achieved and demonstrated in respect of remotely piloted missions. Clearly, as the opening paragraphs of this article demonstrate, there are occasions when errors are made, but the making of errors does not call into question the lawfulness of the method of warfare as such. Rather, it is whether the method is capable of being employed in accordance with established legal requirements that is the critical issue under weapons law.Footnote 38
As the previous section made clear, in certain narrowly defined generic circumstances autonomous attacks are also capable of being conducted in accordance with the requirements of the law of armed conflict. In the cyber domain, however, much will depend on the particular cyber tool that it is planned to use, on the characteristics of that tool, on whether the damaging effect of the cyber tool can be reasonably limited to the intended target of attack, and on whether enough is known about the target computer system to enable proper precautionary judgements of the sort discussed above to be made.
API requires that ‘in the study, development, acquisition or adoption of a new weapon, means or method of warfare, a High Contracting Party is under an obligation to determine whether its employment would, in some or all circumstances, be prohibited by [API] or by any other rule of international law applicable to the High Contracting Party’.Footnote 39 Having concluded that cyber capabilities that are to be used to cause death, injury, damage or destruction to an opposing party to a conflict are means of warfare for the purposes of Article 36, it is clear that a legal review of such capabilities will be required and that the matters discussed in the previous paragraph will need to be considered when deciding whether the capability is indiscriminate by nature.Footnote 40
Liability considerations
Liability for error in remote attack
Legal discussion of remote attack technologies often centres on the question of responsibility. Who is responsible when something goes wrong? In the case of cyber attacks it may be very difficult to determine who precisely undertook the attack and with what particular purpose. The computer from which the attack was initiated may in some cases be identifiable, but the name of the person who created the cyber weapon, the name of the potentially different person who sent the cyber weapon on its way and the state, group, or other entity for which these persons were acting may never be known or capable of public disclosure. These difficulties may therefore make it impossible in practice to fix liability in the case of particular cyber events.
Responsibility, and the related notion of liability, can arise in differing contexts, including at the political/diplomatic level, in the media, at international law, and in domestic law. It may take the form of individual, including command, or state responsibility.
Media coverage of an incident may inform, or drive, perceived political responsibility for the event, as indeed political appreciations may influence media coverage. Early media reports, which may be based in whole or in part on flawed information, speculation, and assumption, and the responses thereto, may fix in the public mind a perception of responsibility that may be hard later to dispel if more reliable data come to light. Early disclosure by governments of factual data, including imagery, may be critical here. This implies, in policy terms, a need to have relevant information readily available in disclosable form if states are to engage successfully in the modern information and media campaigns. Responsibility tends to be attributed by the media to states, but if evidence of individual wrongdoing emerges within the period of active press interest the relevant persons may also attract critical media comment.
When it comes to attributing legal responsibility, judgements after the event must be based on the information, from all sources, that was reasonably available to the decision-maker at the relevant time.Footnote 41 In the case of an attack using a remotely piloted vehicle, the decision by the platform controller to undertake that attack will have been informed by the data fed to him when he was considering and making that decision. The vital issue will be whether that controller's decision to attack was reasonable in the circumstances as they were presented to him. Relevant questions may include whether there were any additional practicable precautions that were not taken and that, if taken, would have verified the status of the target as a military objective, whether the attack could be expected to be proportionate and whether it was being undertaken so as to minimize civilian injury and damage.Footnote 42
It follows that if the relevant equipment was operating properly,Footnote 43 the operator of the platform is liable for his actions in relation to that platform. However, if for example the data feeds to the controller were adversely affected by a system fault, and if that fault can properly be said to have caused the erroneous decision to attack, then the system failure is likely to exonerate the controller from responsibility for the attack.
Similarly, if the opposing party to the conflict, whether through ruses, perfidy, voluntary or involuntary human-shielding or otherwise, materially impedes the platform operator's task, that will also be a factor to take into account when determining responsibility for the resulting events. It would not seem to be reasonable to lay blame at the door of the operator for errors attributable to the supporting systems, enemy action or other causes beyond his control. Whether the erroneous attack truly was beyond the operator's control will, however, be a question of fact to be assessed when all relevant information is available. It would seem that the factors to consider when determining potential liability of the controller of a remote platform are essentially similar to those that apply, for example, in the case of a pilot undertaking a similar mission.
There is no war crime of failing to take precautions in attack. Relevant war crimes under the Rome Statute, for example, would include directing attacks at civilians,Footnote 44 directing attacks at civilian objectsFootnote 45 and prosecuting disproportionate attacks.Footnote 46 The intent that is an ingredient of these offences is not of course to be equated with a failure to take the required precautions, although in particular factual circumstances such a failure may be an element in such an intentional attack. Command responsibility would also be determined on a similar basis to that applying in relation to more conventional military operations, for example bombardment from piloted aircraft. A military commander is criminally responsible under the Rome Statute for crimes committed by forces under his or her effective command and control as a result of his or her failure to exercise control properly over such forces. The provision requires that either the military commander knew, or in the circumstances at the time should have known, that the forces were committing or about to commit such crimes and that he or she failed to take ‘all necessary and reasonable measures within his or her power to repress or prevent their commission’.Footnote 47 While the failure being discussed in the present article, namely the failure to take adequate precautions, does not amount to a war crime under the Rome Statute, any argument that commanders are also responsible for the failure is likely to be assessed according to similar criteria. Ultimately, the issue will be whether the commander knew, or ought to have known, that the method of attack being adopted precluded taking required precautions. It seems most likely that commanders would be aware of this.
Liability for lawful attacks
Generally speaking there is no liability at law for action by the armed forces of one party to an international armed conflict that lawfully causes death, injury, damage or destruction to an opposing party to the conflict.Footnote 48 To be lawful, such action must comply with the law of international armed conflict. Thus there is no liability for the damage lawfully done to military objectives, for the death or injury lawfully caused to members of the opposing armed forces, for expected death, injury or damage to civilians or civilian objects which is not excessive in relation to the anticipated concrete and direct military advantage, or for the death or injury of civilians or damage to civilian objects caused by mistaken or erroneous attacks caused, for example, by the malfunction of military equipment.
The liability to compensate provided in Article 3 of Hague Convention IV, 1907Footnote 49 is repeated in similar terms in Article 91 of API.Footnote 50 Applying Article 91, it would therefore seem that if, as a result of the failure to take all feasible precautions in relation to a remote attack operation, the attack causes excessive death or injury to civilians or excessive damage or destruction to civilian objects in relation to the concrete and direct military advantage anticipated there is likely to be a legal liability to compensate the affected civilians or civilian institutions if the case so demands. The API Commentary suggests that a simple violation of the law of armed conflict is not sufficient, that there must have been loss or damage and that compensation will only be appropriate if restitution in kind or the restoration of the pre-existing position is not possible.Footnote 51 This would suggest that, in order to establish liability, the claimants would need to prove that legally required precautions were not taken,Footnote 52 that the claimants have suffered loss meriting the award of compensation, that this loss was caused by the failure to take precautionsFootnote 53 and that the case demands the award of compensation.
If the injury to civilians and/or damage to civilian objects was caused by a technical malfunction of the equipment, such as faulty software, a manufacturing defect or the erroneous insertion of data during mission preparation, complex issues are likely to confront any attempt to ascribe individual responsibility. Military personnel who act negligently will be subject to their military discipline code, while available action against negligent civilians will depend on their employment contract. If, however, the error that has occurred is such that the incident cannot properly be described as a violation, the law of armed conflict will not require the payment of compensation.Footnote 54 Specifically, it would seem difficult to characterize the negligent manufacture of weaponry as a violation such as to form the basis for a possible claim for compensation under Article 91.Footnote 55 Whether in a particular case a claim would lie under product liability law would depend on the terms of the particular legislation of the relevant state and on the ability of the claimants to bring the claim within the jurisdiction of that state's civil law courts. Such issues lie outside the scope of the present article.
Does remote attack amount to a legally significant change in the conduct of warfare?
Remoteness of attack would be legally significant were it to render rules of targeting inoperable, or to render it impossible to allocate criminal responsibility for wrongful acts or to adjudge whether compensation is payable for attacks that have unsatisfactory consequences.
There are, as we have seen, kinds of remote attack that do not pose such challenges. Thus, when a remotely piloted aerial vehicle is used to attack a target, the role of the remote pilot, usually referred to as the operator, mirrors that of a pilot of a manned aircraft such that targeting law rules can be applied in the same or a similar way, such that criminal liability could lie against the operator, say, in respect of a deliberate attack on civilians and compensation liability could be assessed and decided upon as in the case of an attack using a manned platform.
Moreover, in a sense, man has sought to fight from a distance since the earliest times. Concerns as to the ethics of such developments also date from ancient history.Footnote 56 The trebuchet, cannon, crossbow and longbow, artillery, bombardment from the air, and remotely piloted UAVs can all be regarded as technologically more refined methods of delivering offensive force against the enemy while incurring relatively less risk for one's own forces. This notion of seeking to protect oneself while placing the enemy at enhanced risk is of course central to many methods of warfare, which suggests that remoteness of the operator, per se, does not constitute a qualitative, and thus legally significant, change from what has gone before.Footnote 57 Perhaps the common thread here is that responsibility for attack decisions could always be readily ascribed at the personal, command and national levels. There will frequently be complications, for example where personnel from one nation on detached duty undertake attacks using platforms belonging to a state other than their own, either within a coalition or otherwise;Footnote 58 but those complications do not alter the fact that the person who ordered the attack, and the individuals who carried it out, can be identified and thus responsibility in the senses discussed in this article can be ascribed. Increasing the distance between the attacking individual and the scene where the destruction occurs does not, of itself, seem to change that. Rather, the issue seems to have more to do not so much with distance as with depersonalization altogether.
The anonymity or potential anonymity of a cyber attacker, the impossibility for the affected party to establish whose wrongful act caused an autonomous platform, say, to attack a civilian compound instead of a military objective, are examples of the sorts of circumstances in which we can say that these forms of remote attack would be starting to pose challenges for the law of targeting.
So let us consider autonomous attack technology a little further. If the platform belongs to and is operated by the armed forces of a state, that state will, it is suggested, have similar responsibility for what that piece of equipment does in the battle space to its responsibility for the death, injury or damage caused, for example, by a missile or bomb fired using more conventional, manned technology. In other words, Article 91 of API will determine whether there is a legal obligation to compensate, and the state will retain the discretion whether to make an ex gratia payment in circumstances where no legal liability can be, or has been, established.
Some may seek to conclude from this that if, for whatever reason, a platform autonomously decides to make civilians or civilian objects the object of attack that would prima facie constitute a breach of, respectively, Articles 51(2) or 52(1) of API and would thus constitute a violation for the purposes of Article 91. The alternative view, which the author prefers, would take into account the design of the controlling software, the data fed into the mission control equipment, the settings applied to the algorithm-based technology, and any other information that would demonstrate what the persons planning and commanding the mission intended that the machine should attack. According to this alternative view, the ‘object’ of an autonomous attack consists of the object(s) and/or person(s) that the target recognition equipment was designed or intended to engage. According to this latter view, the machine is using its autonomous capability to achieve the object, or purpose, set for it by those individuals in charge of the mission, with the implication that liability to compensate will only be established under Article 91 if it can be shown that those planners and commanders had as their object of attack the protected persons or objects.
Where personal responsibility for erroneous autonomous attack is concerned, it would seem sensible to conclude that individuals will generally be responsible for their own actions in relation to the autonomous platform, its navigation, and its offensive operation.Footnote 59 If an individual were deliberately to configure the autonomous target acquisition software with the intention that the platform would target civilians and/or civilian objects, it follows that that would amount to a war crime in just the same way as using conventional capabilities with a similar intent would be.Footnote 60 If a failure to take required precautions, however, causes an erroneous autonomous attack a war crime is unlikely to be established; compensation may be payable if the requirements for establishing liability under Article 91 can be established; and individuals responsible for the failure to take precautions may be disciplined, for example on the basis of negligent performance of duties, to the extent this is provided for in applicable armed forces discipline legislation or in the contract of civilian employment.
Conclusion
The tentative conclusion that emerges from this discussion is that the established framework, whether in respect of war crimes, liability to compensate or domestic armed forces or civilian employment discipline, should be capable of being applied, and therefore ought in fact to be applied, in the event of erroneous autonomous attacks. Persons who, in an international armed conflict, use autonomous technology deliberately to undertake unlawful attacks thereby breach the law of armed conflict as do those who use more conventional weaponry to like purpose. The fact that a machine is designed to act autonomously does not absolve those who give orders for the mission, those who plan the mission and those who take the necessary steps to enable the mission to be undertaken of responsibility for their own actions, and it is in the actions of those individuals that the basis for any criminality and liability to compensate is likely to be found.
Suggestions that criminal proceedings be taken against the machine are currently grounded in fiction. However, as notions of artificial intelligence (AI) continue to mature, it is conceivable that a point will arise at which human involvement is so remote, in a causal sense, from the decision to attack that commanders and planners can no longer sensibly be held accountable. In the author's view, we have not got to that point yet, but as technology becomes more complex and as decision-making relies increasingly on AI and less and less on human perception and judgement, the focus for responsibility may be expected to shift from planners and commanders to software engineers and the robots they beget.
